Camera time interval topology network construction method based on forgetful neural network algorithm

Abstract

The application discloses a camera time-distance topology network construction method based on a forgetting neural network algorithm, and comprises the following specific steps: S1, marking each camera as a point in a camera topology network; S2, drawing a directed edge in the camera topology network according to a pedestrian action track recorded by the camera; and S3, a server analyzes and judges whether a pedestrian collected by the camera appears for the first time, updates the weight of the directed edge in the camera topology network, and completes the construction of the topology network. The application can provide the relative distance between each camera on the time level, is used for assisting personnel unique identification, greatly improves the accuracy of face recognition, and provides an important technical basis for personnel passage design and escape passage design.

Description

Technical Field

[0001] This invention relates to the field of time-distance analysis algorithms between cameras, and in particular to a method for constructing a camera time-distance topology network based on a forgetting neural network algorithm. Background Technology

[0002] In reality, after network cameras are deployed, only their IP addresses and the geographical location information transmitted by users are known. Their actual effective distance information cannot be effectively obtained, making it impossible to determine the relative positions of different cameras in reality. This poses significant challenges for tasks such as designing pedestrian passageways, fire escape routes, and security checks. Furthermore, due to the lack of effective temporal distance relationships, it is difficult to accurately identify individuals in facial recognition processes by considering the spatiotemporal relationship of their presence. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing camera temporal topology networks based on the forgetting neural network algorithm.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] The method for constructing a camera temporal topology network based on the forgetting neural network algorithm includes the following specific steps:

[0006] S1: Mark each camera as a point in the camera topology network;

[0007] S2: Draw the directed edges in the camera topology network based on the pedestrian movement trajectories recorded by the cameras;

[0008] S3: The server analyzes and determines whether the pedestrian captured by the camera is appearing for the first time, updates the weights of the directed edges in the camera topology network, and completes the construction of the topology network.

[0009] Specifically, S2 involves connecting the cameras passed by each pedestrian in sequence according to the time order of their passing, forming multiple edges with vector characteristics.

[0010] The weights in S3 include time interval weight T and memory weight M. The time interval weight represents the time interval between two points, and the memory weight represents the reliability of the time interval.

[0011] Specifically, S3 is:

[0012] S301: Extract human features using pedestrian recognition technology and facial recognition technology to determine the identity of pedestrians passing by the camera;

[0013] S302: Record pedestrian identities into the database based on feature values, and record the last position of each pedestrian's last appearance at the camera, denoted as LP;

[0014] S303: The server searches the database for the location LP where pedestrian P last appeared on the camera. P Determine whether pedestrian P is appearing for the first time;

[0015] S304: If pedestrian P is appearing for the first time, then LP P Records are sent back to the database; if pedestrian P is not appearing for the first time, the weights of the directed edges of the camera topology network are updated based on the shooting time difference between the previous appearance point and the current appearance point, i.e., the possible time interval.

[0016] Specifically, S304 is:

[0017] Once camera B notices pedestrian P and extracts the pedestrian's feature values, it accesses the server to search for LP in the database. P If not found, then P is the first occurrence, and camera B is taken as LP. P The record is sent back to the database; if the LP is found... P Given camera A, the time difference between the shots taken by A and B is a possible time interval. The camera time interval topology network is updated based on the possible time intervals.

[0018] If it exists Then update the double weights on the edges;

[0019] If it does not exist Then establish directed edges And assign it an initial double weight value.

[0020] The weights of the directed edges in the updated camera topology network are specifically as follows:

[0021] When the edge During the initial construction, the possible time interval Δt0 between the current and previous images of the pedestrian is taken as the initial time interval weight.

[0022] When the edge When a new possible time interval Δt1 is received after the structure is constructed, Change value for:

[0023]

[0024] Where f(x) and h(x) are monotonically decreasing functions defined in (0, +∞) with a range of [1, 0); μ is a constant representing the step size;

[0025] When no time interval can be passed in, the memory weight M will decrease by the value of dm every dt seconds, where

[0026] dm = g(x)dt;

[0027] When a possible time interval is passed in, the memory weight M will increase by a constant φ.

[0028] The method for constructing a camera temporal topology network based on the forgetting neural network algorithm also includes the following steps: storing the data of the camera temporal topology network, including centralized storage and distributed storage;

[0029] When data is stored centrally: all camera time-range topology information is stored in the central data server and backed up using a backup data server;

[0030] When data is stored in a distributed manner: the data is stored in cameras with computing power, edge servers, or AI NVRs.

[0031] The beneficial effects of this invention are:

[0032] This invention can provide the relative distance between each camera in time, which can be used to assist in the unique identification of personnel, greatly improve the accuracy of facial recognition, and provide an important technical foundation for the design of personnel passages and escape routes. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0034] Figure 1 This is a flowchart of the present invention;

[0035] Figure 2 This is a flowchart of the centralized storage time-domain topology network of the present invention;

[0036] Figure 3 This is a flowchart of the distributed storage time-domain topology network of the present invention. Detailed Implementation

[0037] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0039] Explanation of the concept of time interval: If a person P appears in camera A at time t0, and then appears in camera B at time t1, and does not appear in any other camera between t0 and t1, then Δ t =t1-t0 is a possible time interval between cameras A and B. After a long period of time, all possible time intervals are counted, and the most likely value of Δt in most cases is calculated. This value is taken as the time distance between A and B, called the time interval from A to B, denoted as Δt. (Time Distance). It's easy to see that the distance between cameras is measured by the time it takes to reach each other. Even if cameras are tens of thousands of miles apart, as long as most people can pass by the cameras in a short time, their time distance is still very small. Here are two special cases: 1. The time distance between cameras deployed at airports in two cities could be very small or very large. There might be multiple time zones between the two cameras, each with its own frequency of occurrence. Mathematically, this algorithm will ultimately select the time zone with the highest frequency. 2. The time it takes for a person to travel from A to B and from B to A might be different. This means there is a two-way time distance inconsistency, i.e.,...

[0040] Explanation of camera topology: In the construction of the camera topology, cameras are abstracted as points, and these points are connected by directed edges. The time-distance weight of an edge represents the time distance between two points. If there is no edge between two points, there are two possibilities: 1. There was once an edge between them, but it was deleted because the frequency of human traversal was too low; 2. The aforementioned Δt has not yet been collected between them.

[0041] Explanation of directed edges in a camera topology: In a topology network, a camera is considered a point, and the line connecting any two points is an edge. This edge has vector characteristics, meaning it has direction and length. In this project, this vector includes two levels of attributes: temporal weight and memory weight. The temporal weight represents the time interval between two points (cameras), while the memory weight represents the reliability of the temporal interval. The more pedestrians sequentially passing between two points in a recent period, the more reliable the temporal information, and thus the larger the memory weight. A directed edge from A to B is denoted as... but The time interval weight is denoted as (time), whose memory weight is denoted as (memory), both are greater than 0.

[0042] Modeling methods for camera time-distance topological graph (CTTG).

[0043] I. Construction of CTTG midpoints

[0044] First, all cameras are searched and treated as their corresponding points in CTTG. Then, edges between the points are drawn based on the camera communication content.

[0045] II. Edge Construction in CTTG

[0046] Cameras use pedestrian recognition and facial recognition technologies to extract human feature values ​​to determine whether pedestrians in two images belong to the same person. In addition to the database recording pedestrian identities based on feature values, an extra piece of information needs to be recorded: which camera in the CTTG system each pedestrian last appeared in. This record of the last position is called the last position: LP, where LP is denoted as LP for pedestrian n. n .

[0047] After camera B notices pedestrian P and extracts the pedestrian's feature values, it accesses the server and searches for LP in the database using certain methods. P If not found, then P is the first occurrence, and camera B is taken as LP. P The record is sent back to the database; if the LP is found... P And its value is camera A. The algorithm will calculate the time difference between the shooting times of A and B (i.e., the possible time interval). This is called the possible time interval input. CTTG will update itself based on the possible time interval. At this time, there are two cases: 1. If there exists Then update the double weights on the edge; 2. If it does not exist Then establish directed edges And assign it an initial double weight value.

[0048] 1. Edge weight update method and initial value

[0049] 1) Time interval weight T

[0050] When the edge During the initial construction, the possible time interval Δt0 between the current and previous images of the pedestrian was taken as... This is The initial value.

[0051] if After the initial construction, a new possible time interval (AB again sequentially photographs a certain person) Δt1 is received. Change value for

[0052]

[0053] Here, f(x) and h(x) are monotonically decreasing functions defined on the domain (0, +∞) and have a range of (0, 1).

[0054] μ is the update step size, a constant defined by the user.

[0055] In simple terms, the above formula means: the algorithm trusts the original data more. The value, so when the new time interval Δt1 is received, if it is the same as... The greater the difference, the better. The smaller the change in value, the higher the weight representing reliability. The larger the value, the more the algorithm trusts the original data. value, The smaller the change, the better. This is an algorithm based on empiricism.

[0056] 2) Memory weight M

[0057] Memory weights are affected by two factors: 1. the passage of time; 2. the introduction of new possible time intervals.

[0058] Time passes:

[0059] When no time interval can be passed in, the memory weight M will decrease by the value of dm every dt seconds, where

[0060] dm=g(m)dt

[0061] The function g(x) must satisfy the following condition:

[0062] 1. Its domain is [0, e], where e is a constant depending on the situation, and the value M of the memory weight will not be outside [0, e].

[0063] 2.g(0)=0, g(e)=0, g(x)<0.

[0064] 3. There are constants b and c, depending on the case, located in [0, e]. Where b is greater than c and g(b) is the minimum value of the function g(x).

[0065] 4. The function is continuous on its domain, monotonically decreasing on [0, b], and monotonically increasing on [b, e].

[0066] Possible time interval input:

[0067] When a possible time interval is passed in, the memory weight M increases by a constant φ. This φ is a positive constant, depending on the situation, and φ <e。

[0068] If M + φ is greater than e, then force M = e - α, where α is also a constant depending on the situation, and α < e. <e。

[0069] In simple terms, M is a value with an upper limit of e that decreases continuously over time, and the rate at which it decreases is related to the size of M itself. When a new possible time interval is introduced, M will increase by a constant. If M increases to a size greater than e, then M is forced to equal a value that is only slightly smaller than e, and then M continues to decrease over time.

[0070] When the value of M is less than c in g(x), the algorithm will consider that no one has passed through this edge for too long, and will delete this edge and its record. This is called the algorithm forgetting this edge.

[0071] When the server storage space of the storage model is insufficient, it will select several edges with the lowest forgetting memory weight to free up storage space.

[0072] III. CTTG Storage Method

[0073] 1. When data is centrally stored: There is a central data server and a backup data server. All camera time-range topology information is stored in the central data server and backed up using the backup data server.

[0074] 2. When data is stored in a distributed manner: that is, when data is stored in cameras with computing power, edge servers, or AI NVRs, then each camera can be treated as a point (if it is an AI NVR or edge server, it remembers the cameras it hosts), remembering the edges originating from it, as well as the other points (other cameras) connected to it, for example, the edges. The storage method is for camera A to remember the IP address (or addressing method) and edge information of camera B. The two weights.

[0075] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for constructing a camera temporal topology network based on a forgetting neural network algorithm, characterized in that, The specific steps include the following: S1: Mark each camera as a point in the camera topology network; S2: Draw the directed edges in the camera topology network based on the pedestrian movement trajectories recorded by the cameras; S3: The server analyzes and determines whether the pedestrian captured by the camera is appearing for the first time, updates the weights of the directed edges in the camera topology network, and completes the construction of the topology network; specifically: S301: Extract human features using pedestrian recognition technology and facial recognition technology to determine the identity of pedestrians passing by the camera; S302: Record pedestrian identities into the database based on feature values, and record the last position of each pedestrian's last appearance at the camera, denoted as LP; S303: The server searches the database for the location LP where pedestrian P last appeared on the camera. P Determine whether pedestrian P is appearing for the first time; S304: If pedestrian P is appearing for the first time, then LP P Records are sent back to the database; If pedestrian P is not appearing for the first time, then the weights of the directed edges of the camera topology network are updated by the shooting time difference between the previous appearance point and the current appearance point, i.e., the possible time interval. The weights of the directed edges in the updated camera topology network are specifically as follows: When the edge During the initial setup, the possible time interval between the current and previous captures of pedestrians was taken. As initial time interval weight ; When the edge After it was constructed, new possible time intervals emerged. hour, Change value for: ， in and The domain is The function is a monotonically decreasing function with a range of . ; The constant represents the step size; When no time interval can be passed in, the memory weight M will decrease by the value of dm every dt seconds, where ； Indicates when Function of time ;function In order to be in Continuous on the domain, Monotonically decreasing above, in The function is monotonically increasing, and , b and c are located at The constant in, where b is greater than c and For function The minimum value; When a possible time interval is passed in, the memory weight M will increase by a constant φ, and φ < 0. If M+φ> Then force M= , ; The forgetting neural network algorithm states that: when the value of the memory weight M is less than the function... When the function value of c in the model is less than or equal to 0, the algorithm deletes the edge and its record, indicating that the algorithm forgets the edge. When the server storage space of the storage model is insufficient, the algorithm selects to forget the edges with the smallest memory weight M to free up storage space.

2. The method for constructing a camera temporal topology network based on the forgetting neural network algorithm according to claim 1, characterized in that, Specifically, S2 involves connecting the cameras passed by each pedestrian in sequence according to the time order of their passing, forming multiple edges with vector characteristics.

3. The method for constructing a camera temporal topology network based on the forgetting neural network algorithm according to claim 1, characterized in that, The weights in S3 include time interval weight T and memory weight M. The time interval weight represents the time interval between two points, and the memory weight represents the reliability of the time interval.

4. The method for constructing a camera temporal topology network based on the forgetting neural network algorithm according to claim 1, characterized in that, Specifically, S304 is: Once camera B notices pedestrian P and extracts the pedestrian's feature values, it accesses the server to search for LP in the database. P If not found, then P is the first occurrence, and camera B is taken as LP. P The record is sent back to the database; if the LP is found... P Given camera A, the time difference between the shots taken by A and B is a possible time interval. The camera time interval topology network is updated based on the possible time intervals. If it exists If so, then update the double weights on the edge; If it does not exist Then a directed edge is established. And assign it an initial double weight value.

5. The method for constructing a camera temporal topology network based on the forgetting neural network algorithm according to claim 1, characterized in that, It also includes the following steps: Data from the camera time-range topology network is stored, including centralized storage and distributed storage; When data is stored centrally: all camera time-range topology information is stored in the central data server and backed up using a backup data server; When data is stored in a distributed manner: the data is stored in cameras with computing power, edge servers, or AI NVRs.

Images